Multimodal LLMs are not all you need for Pediatric Speech Language Pathology

Problem
This preprint addresses the significant gap in the application of multimodal large language models (LLMs) for Pediatric Speech Language Pathology (SLP), specifically in the context of Speech Sound Disorders (SSD). SSD affects approximately 5% of children, and the existing literature highlights the challenges faced by speech-language pathologists, including staffing shortages and overwhelming caseloads. The authors argue that while LLMs have shown promise in various domains, they are not sufficient for the nuanced classification tasks required in pediatric SLP, particularly when dealing with the complexities of SSD.

Method
The authors propose a hierarchical classification framework that cascades from binary classification to more granular type and symptom classification. This approach leverages Speech Representation Models (SRMs) fine-tuned on the SLPHelmUltraSuitePlus benchmark, which encompasses multiple clinical tasks. The methodology includes targeted data augmentation techniques designed to address biases identified in prior research. The authors also extend their data augmentation strategy to Automatic Speech Recognition (ASR) tasks, enhancing the robustness of the model. The training compute details are not disclosed, but the focus on fine-tuning SRMs suggests a significant computational investment in optimizing model performance across the benchmark tasks.

Results
The results indicate that the proposed SRM-based approach significantly outperforms the state-of-the-art LLM-based methods across all evaluated tasks in the SLPHelmUltraSuitePlus benchmark. Specific performance metrics are not detailed in the abstract, but the authors claim “large margin” improvements, suggesting effect sizes that are substantial enough to warrant attention. The results underscore the efficacy of the hierarchical classification strategy and the targeted data augmentation in improving SSD classification accuracy.

Limitations
The authors acknowledge several limitations, including the potential for overfitting due to the limited size of the training dataset and the need for further validation across diverse populations. They also note that while their approach improves upon LLMs, it may not fully generalize to all aspects of pediatric SLP, particularly in real-world clinical settings. An additional limitation not explicitly mentioned is the reliance on SRMs, which may not capture the full multimodal context that could be beneficial in SLP applications.

Why it matters
This work has significant implications for the field of pediatric SLP, as it provides a more effective alternative to LLMs for SSD classification tasks. By demonstrating the superiority of SRMs in this context, the authors pave the way for future research that could further refine these models and explore their application in clinical settings. The publication of their models and code encourages reproducibility and further exploration of hierarchical approaches in SLP, potentially leading to improved diagnostic tools and interventions for children with speech sound disorders.

Authors: Darren Fürst, Sebastian Steindl, Ulrich Schäfer
Source: arXiv:2604.26568
URL: https://arxiv.org/abs/2604.26568v1